OpenJarvis: Stanford's Local-First Framework for Building On-Device AI Agents
Stanford researchers from the Hazy Research and Scaling Intelligence Lab have released OpenJarvis, an open-source framework for building personal AI agents that run entirely on-device.
OpenJarvis defines five composable primitives — Intelligence, Engine, Agents, Tools & Memory, and Learning — replacing ad-hoc integrations with a structured, opinionated stack. The framework prioritizes efficiency, treating energy, FLOPs, latency, and cost as first-class constraints alongside task quality. According to the team's research, local language models already handle 88.7% of single-turn queries, with efficiency improving 5.3x from 2023 to 2025.
Developer interfaces include a browser app, desktop applications for macOS/Windows/Linux, a Python SDK, and a CLI. The codebase is primarily Python (77.6%) with performance-critical Rust components (14.4%). Sponsors include Google Cloud, Ollama, and IBM Research.
The project has 1,000 GitHub stars with active development since its February 2026 creation.
https://github.com/open-jarvis/OpenJarvis
← Back to all articles
OpenJarvis defines five composable primitives — Intelligence, Engine, Agents, Tools & Memory, and Learning — replacing ad-hoc integrations with a structured, opinionated stack. The framework prioritizes efficiency, treating energy, FLOPs, latency, and cost as first-class constraints alongside task quality. According to the team's research, local language models already handle 88.7% of single-turn queries, with efficiency improving 5.3x from 2023 to 2025.
Developer interfaces include a browser app, desktop applications for macOS/Windows/Linux, a Python SDK, and a CLI. The codebase is primarily Python (77.6%) with performance-critical Rust components (14.4%). Sponsors include Google Cloud, Ollama, and IBM Research.
The project has 1,000 GitHub stars with active development since its February 2026 creation.
https://github.com/open-jarvis/OpenJarvis